Components

Generate Text (SGLang)

@lmsys

Drives a chat turn against a colocated SGLang service from a user and system String, streaming the reply as String chunks and closing with a StreamEnd; supports JSON-object output and rolling conversation history.

Ground Phrase (Florence 2)

@microsoft

Localises the noun phrases of a caption in an image: it runs Florence-2 phrase grounding over an Image and a String, emitting either bare [Rectangle] boxes or labelled [{bounds, label}] regions depending on how the output is wired.

Send Object Counts (MQTT)

Terminal MQTT sink that publishes per-class object counts as Cumulocity SmartREST measurements. It takes [(String, UInt64)] class-name and count pairs, reports each count plus a total, and raises per-class and overall alarms when any count reaches the configured limit.

Input Browser Webcam

Receives a browser's camera and microphone capture over WebRTC and emits a VideoFrame stream and an AudioFrame stream, each gated on independently, so the graph can take live media from a remote browser tab.

Detect Face Liveness (MiniFASNet)

@deepinsight

Scores each Rectangle face box in an Image for passive presentation-attack detection with a dual MiniFASNet anti-spoofing ensemble, emitting one liveness Classification per face -- live / print / replay plus confidence -- in box order.

Classify Text (SetFit)

@huggingface

Few-shot SetFit classifier that scores a String against a class set fitted from a handful of examples, emitting ranked [Classification], a single Classification, or [Double] in sorted class order. For few examples per class with no LLM endpoint.

Describe Image Regions (Florence 2)

@microsoft

Florence-2 dense region captioning: takes an Image, proposes regions, and emits [{bounds: Rectangle, caption: String}] — one short free-text caption per region, in pixel space, with no class id or confidence.

Detect Objects (Ultralytics YOLO)

@ultralytics

Runs a bound Ultralytics YOLO checkpoint over each Image with NMS and projects the detections onto the connected output shape: boxes, geometry, classes, masks, a count, or an overlay. The mask shapes need YOLO-seg weights.

Detect Faces (Auraface)

@fal

SCRFD face detector over an Image that emits [BoundingBox] or [(BoundingBox, [Landmark])] — one box plus optional five-point facial keypoints per face — with the arm chosen by the downstream connection. Use to find and align faces.

Upscale Image (DRCT)

@ming053l

Super-resolves an Image with the DRCT transformer at a fixed integer factor matched to the checkpoint, optionally tiling large frames to bound VRAM, and emits the enlarged Image.

Remove Document Shadow (DocShadow)

@cxh_research

DocShadow (SD7K) document shadow remover. Runs each shadow-cast Image through a learned restoration model that separates uneven illumination from text detail and emits an evenly-lit Image. Tiny inputs pass through unchanged.

Classify Speech Emotion (HuggingFace)

@huggingface

Runs any HuggingFace audio-classification checkpoint for speech affect over each AudioFrame, emitting the ranked emotion label/score pairs as [Classification] or only the single top emotion as a String.

Estimate Optical Flow (Classical)

@opencv

CPU dense optical flow between Image frames using a selectable classical method (method = DIS, Farnebäck, or TV-L1), emitting one OpticalFlow (H, W, 2) tensor per frame at the source resolution, with no model service or GPU required.

Check Geometry Within ROI

@bleenco

Spatial containment test: takes a geometry item (a box, point, mask, polygon or point list) and an ROI shape, and emits a Bool indicating whether the item lies inside or outside the region under the configured rule.

Animate Portrait Audio (Sonic)

@tencent

Audio-driven talking head with Sonic: consumes a driving AudioFrame stream and animates a bound source portrait through a Stable-Video-Diffusion backbone so lip motion and head movement follow the audio, emitting the result as a sequence of Image frames.

Constant

Pure source with no input port: emits a fixed deploy-time literal forever, re-sending the same value on every downstream demand. The configured value carries its own type tag, so it pins the output stream type t to a concrete type rather than leaving it free.

Output Point Cloud File

Writes each upstream PointCloud to a file under output_path in the layout chosen by file_format (binary or ASCII PLY, PCD, or XYZ), subsampling with every_n and bounding retention with max_files.

Segment Image (SAM 3)

@meta

Concept-prompted open-vocabulary segmentation on an Image. Each noun phrase is segmented and the sink resolves the projection: mask-plus-box pairs, masks, boxes, parallel arrays, a merged Mask, a UInt64 count, or a colourised label Image.